Reader question: How can a small team use AI to compare two long documents without turning a convenient summary into the evidence?

On September 10, 2026, current product guidance names document comparison as a use case. OpenAI lists comparing two documents and applying a framework from one to another. Gemini Notebook lets users select sources and open citations at quoted locations. Adobe documents citations identifying a file, section, and page. OpenAI’s file-upload guidance Gemini Notebook’s chat guidance Adobe’s citation guidance

These are capability descriptions, not proof of a complete comparison. The useful goal is a difference ledger with source locations and unanswered questions.

Define the comparison before uploading anything

“Compare these” is too vague for consequential work. Name the difference that matters. A layout change and a changed cancellation term should not receive the same treatment.

Write one sentence that names the question and boundary: “Find changes affecting eligibility, fees, and deadlines between Document A, dated May 1, and Document B, dated September 1. Do not recommend which version to adopt.” Then record:

  • The decision and owner: what may be decided, and who can resolve an unresolved difference.
  • The documents: exact titles, versions, dates, and owners.
  • The scope: included topics and deliberate exclusions.
  • The evidence format: document label, page or section, and a short passage for every substantive row.

Use the team’s approved service and workspace for the material. A long file can still contain personal information, confidential terms, or access-restricted content. A comparison prompt does not grant permission to upload it.

Create a page-aware source register

Prepare the two files before asking for a comparison. Save an unchanged copy of each, record the filename and version date, and note whether pages are searchable text, scanned images, or a mixture. Add a file fingerprint if available.

Make a source register with document ID, file title, version date, total pages, printed-page offset, section map, visual pages, and access owner. A PDF’s page 1 may contain a cover while the footer begins at page 3, so keep both numbers.

This step catches a common false match. A model may find the same heading in both files while missing a table, footnote, appendix, or scanned signature. OpenAI says that, outside Enterprise visual retrieval for PDFs, document retrieval is text-based and images embedded in documents are discarded. OpenAI’s file-upload limits and retrieval note Mark charts, screenshots, handwritten notes, and scans for visual inspection or separate transcription. Do not let a clean text extraction imply that the whole page was read.

Ask for a difference ledger, not a verdict

Request one row per meaningful difference:

Compare only Document A and Document B. Return one row for each added, removed, changed, unchanged, or unclear item. Include the topic, A and B locations, short evidence passages, classification, and the owner’s unresolved question. Do not fill gaps or infer a missing term. If no source location is available, mark the row unclear.

Keep unchanged rows for high-impact sections, put cosmetic edits separately, and distinguish “not found” from “removed.” Absence in an excerpt is not proof of deletion.

If the tool offers source selection, choose only the two registered files. Google says Gemini Notebook can include or exclude individual sources and uses direct text, images, and quotations from selected sources for citations. Google’s source-selection guidance That bounds the question, but does not make the result authoritative.

Consider a synthetic example. A 70-page employee handbook is compared with a new edition. The first pass reports: changed, leave window, A p. 18 §3.1, B p. 20 §3.1. A footnote in B may limit the change to new hires. The right entry is “changed; application unresolved,” not “the leave window is now 14 days.” This is illustrative, not an executed test.

Verify every consequential row at the source

The comparison output is a map for checking. Open the cited page in both documents. Read surrounding paragraphs, definitions, exceptions, tables, footnotes, and appendices. Confirm a meaning change, not a page shift or renamed heading.

Record exact locations yourself, including PDF page and printed page when needed. If a citation gives a passage but no page, find it in the registered copy and add the page. If the tool points nearby, label the row citation mismatch and recheck it. Do not repair a missing citation with a guess.

Adobe says its citations open highlighted source content, but warns they do not guarantee completeness or correctness. Adobe’s citation review guidance Google says precise questions help retrieval and that an answer not present in selected sources will not be provided. Gemini Notebook FAQ A locator lowers search time; it does not remove the source check.

Hand off the record, then stop when it is unclear

The comparison owner hands over the source register, exact copies, difference ledger, verified high-impact rows, and unresolved rows with an owner and next question. Keep the AI response separate from the authoritative document. The ledger should not overwrite a policy, agreement, or approved brief.

Stop when the edition or page map is unclear, a file is incomplete, a citation does not support the row, a table or image was missed, definitions conflict, or a high-impact difference cannot be resolved. Mark it unresolved, preserve the evidence, and ask the source owner for context.

The best comparison is not the one with the shortest summary. It is the one that lets the next person move from a claim to two pages, see what changed, and tell what still needs a decision. Let AI do the first pass across the long files. Keep the page, the passage, and the uncertainty in the handoff.

Sources and limitations

  • OpenAI Help Center, “File Uploads FAQ” — checked September 10, 2026; supports document comparison, applying a framework from one file to another, file limits, text-based retrieval and the image-handling caveat. Limitation: a supported use case and size limit do not guarantee a complete or accurate comparison; plan, account, and file conditions vary.
  • Google, “Use chat in Gemini Notebook” — checked September 10, 2026; supports source selection, direct source quotations, citation navigation, and the documented need to double-check experimental actions. Limitation: feature availability, citation format, and behavior vary by product, plan, device, and source.
  • Google, “Frequently asked questions” for Gemini Notebook — checked September 10, 2026; supports relevance-based retrieval, precise-question guidance, source limits, and the statement that answers depend on uploaded source material. Limitation: these are product controls, not a guarantee that every relevant passage will be retrieved or interpreted correctly.
  • Adobe Acrobat Help, “View citations in responses” — checked September 10, 2026; supports multi-document citations with document, section, and page metadata and links back to source content. Limitation: the feature is specific to supported Acrobat experiences and formats; citation metadata still requires inspection.
  • Adobe Acrobat Help, “Review citations and source content” — checked September 10, 2026; supports opening highlighted source context and Adobe’s warning that citations do not guarantee completeness or correctness. Limitation: product behavior, access, and source indexing can change.